Dataset: https://cg.cs.tsinghua.edu.cn/traffic-light/ This scripts help the user understand, convert to YOLO format and extract useful information from this dataset.

• get_ds_infos_JSON : Counts each class apparition from "TL_train.json". Useful for understanding the dataset.

• get_image_by_class : Extracts a number of images of a specific class from the images directory. Useful for understanding what each class represents.

  arguments: 1. class_id : id of the selected class 2. number_of_pictures: number of photos samples

• convert_to_yolo : Generates YOLO formatted annotations for the specified class. You should specify the number of annotations you want to extract. Keep in mind that there are cases when exists more than one class. In this case you can chose to select only the class you like ( discard the others), or save the image with all classes.

  This can be challenging when you try to have a balanced dataset. Example: ./convert_to_yolo -c 1 -n 1000

  Will have the following output: [(1, 1000), (2, 723), (3, 574), (4, 198), (5, 309), (6, 417)] This translates as: 1000 samples of the 1 class 723 samples of the 2 class 574 samples of the 3 class ...

  In order to solve this problem you can add "-x 1". This will make the selection more selective. Example: ./convert_to_yolo -c 1 -n 1000 -x 1

  Will have the following output: [(1, 1000), (2, 70), (3, 242), (4, 22), (5, 20), (6, 65)]

  arguments:

  1. -c : selected class id
  2. -n : number of samples
  3. -x : extra priority in selecting the class; use "-x 1" to activate
  4. -exclude : specify the classes you want to exclude. The converted images will not contain any of these classes. Parse them the following way: -exclude 1,2,3 ( will exclude 1, 2, 3 classes) ex: ./convert_to_yolo.py -c 1 -n 1000 -x 1 -exclude 4,5,6 Converts 1000 annotations from class 1 and removes all the annotations from the 4,5,6 classes.

• get_ds_infos : Counts the classes from the converted dataset and prints the information.

  arguments:

  1. -src : directory where from to read the files

• copy_images : Reads annotations from one file and copies the corresponding images with annotations to a one directory. You can parse multiple directory and copy the images and annotations in one folder.

  arguments:

  1. -src : directory from which the files are read; Default is "TL_train_data/home/lyf/develop/data/StreetViewCrops/"
  2. -f : folder paths relative to the current directory
  3. -x : Copies all images in one directory instead of separate ones

1. get_ds_infos_JSON : generates dataset_infos; check the data and see what you have there
2. get_image_by_class : extract some images and see what each class id represent
3. convert_to_yolo : convert some classes from json to txt
4. get_ds_infos : check what you have extracted and make sure your set is equally distributed
5. copy_images : final step, copy the images and annotations in one directory
6. TRAIN THEM!